Dear Nixtla community,

thank you for your amazing work. Is it possible to perform multivaraite time series classificaiton using mlforecast? Eventually, how it could be done?

Thank you very much in advance for support

Dear Nixtla community, I am trying to replicate the example on the AirPassengersDF on another dataset, but I keep getting forecasts (Y_hat_df) with only NaN values except the last one. I am using dates like 2023-10-31 230000, it is the only difference that I can spot with the example. Do you know why and how to fix that? Many thanks

Is there a function like

predict_insample

that you can pass a df to instead of using the last training df? I'd like to evaluate a test set (completely separate unique_ids) by running sliding window forecasts.

Is the ESRNN model in neuralforecast ? It’s mentioned in one of the pages but could not find it in the model list

Trying to make some test in a data frame called panel, that has info in 1 minute time frame that has the next characteristics. Unfortunately Im getting all values a Nan. What could be wrong? How can I fix it? panel.dtypes: unique_id int64 ds datetime64[ns] y float64 Data_Example: unique_id ds y 499995 2024-02-27 05500 1.08497 499996 2024-02-27 05600 1.08495 499997 2024-02-27 05700 1.085 499998 2024-02-27 05800 1.08494 499999 2024-02-27 05900 1.08497 tsfeatures(panel, dict_freqs={'M': 1},features=[acf_features]) unique_id x_acf1 x_acf10 diff1_acf1 diff1_acf10 diff2_acf1 diff2_acf10 0 0 NaN NaN NaN NaN NaN NaN 1 1 NaN NaN NaN NaN NaN NaN 2 2 NaN NaN NaN NaN NaN NaN 3 3 NaN NaN NaN NaN NaN NaN 4 4 NaN NaN NaN NaN NaN NaN ... ... ... ... ... ... ... ... 499995 499995 NaN NaN NaN NaN NaN NaN 499996 499996 NaN NaN NaN NaN NaN NaN 499997 499997 NaN NaN NaN NaN NaN NaN 499998 499998 NaN NaN NaN NaN NaN NaN 499999 499999 NaN NaN NaN NaN NaN NaN

Hi , can someone share python code of data preparation (creation of ds ,y and unique_id) to be used as input for neuralforecast or hierarchical forecasting

has anyone been able to successfully run NeuralForecast on Apple Silicon leveraging GPU acceleration? i cannot get

tensorflow-metal

and

tensorflow

to play nice, even using aarm64 Docker image.

Hi everyone, I'm using TimeGPT to predict a time series with 14 exogenous variables with a frequency of 30 minutes. All exogenous variables have been normalized. Most target values are between 6000-8000. But no matter how I modify the model, the predicted value is always very large, such as: 831199299764224. How should I solve it?

<!channel>, here is our benchmark with SalesForce Morai. Please share some love. https://www.linkedin.com/posts/mergenthaler_happyforecasting-timeseries-forecasting-a[…]340177952768-cMs5?utm_source=share&utm_medium=member_desktop

And here in twitter! Please like and retweet 🫶 https://twitter.com/azulgarza_/status/1773472833840288232

🌟 Exciting News for Forecasting Enthusiasts! 🌟 The latest Kaggle challenge is here, and it’s all about probabilistic forecasting! 🚀 Here is starter notebook using Nixtla statsforecast -> https://www.kaggle.com/code/predaddict/pfi-starter-nixtla-statsforecast https://www.kaggle.com/competitions/probabilistic-forecasting-i-temperature/overview

Hi Team, I am trying to do an evaluation of my forecasts with the

utilsforecast.evaluation.evaluate

function but I keep hitting the following error when adding the train_df parameter. I’ve had a look at the source code which indicates that train_df is optional, however if I omit the train_df I get the same error. call:

from utilsforecast.evaluation import evaluate
from utilsforecast.losses import (
    mse,   # mean square error
    mape,  # mean absolute percentage error
    mae,   # mean absolute error
    mase,  # mean absolute scaled error
    rmse,  # root mean square error
    mqloss, # multi-quantile loss
    scaled_crps, # scaled continues ranked probability score
    )
eval_df = evaluate(forecasts, metrics=[metrics], train_df=train)

> File “/Users/.pyenv/versions/datascience/lib/python3.11/site-packages/utilsforecast/evaluation.py”, line 72, in evaluate > metric_requires_y_train = { > ^ > File “/Users/.pyenv/versions/datascience/lib/python3.11/site-packages/utilsforecast/evaluation.py”, line 73, in dictcomp > _function_name(m): “train_df” in inspect.signature(m).parameters > ^^^^^^^^^^^^^^^^^ > File “/Users/.pyenv/versions/datascience/lib/python3.11/site-packages/utilsforecast/evaluation.py”, line 22, in _function_name > name = f.name > ^^^^^^^^^^ > AttributeError: ‘list’ object has no attribute ‘__name__‘. Did you mean: ‘__ne__‘? my forecasts dataframe has the following columns:

unique_id

ds

Naive

Naive-lo-95

Naive-hi-95

MSTL

MSTL-lo-95

MSTL-hi-95

AutoETS

AutoETS-lo-95

AutoETS-hi-95

combined-lo

combined-hi

combined-base

my train dataframe has the following columns:

ds

y

unique_id

and the metrics list is simply [mse, rmse, mape] Any assistance would be appreciated. Thanks in advance.

Hi Nixtla team. I wanted to know if statsforecast and mlforecast support multivariate forecasting.

I share the new Chronos post with mention of Nixtla and the adjustments they have made compared to the one Nixtla made a couple of weeks ago https://www.linkedin.com/posts/abdulfatir_we-are-elated-to-see-the-overwhelming-response-activity-7181308085615906816-eIUa?utm_source=share&amp;utm_medium=member_android

Hi everyone, I'm trying to use the PredictionIntervals on MLForecast. But I keep getting this error:

mlf.fit(
    train_df,
    prediction_intervals=PredictionIntervals(n_windows=2, h=horizon),
)

Unexpected exception formatting exception. Falling back to standard exception
Traceback (most recent call last):
  File "/opt/conda/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3553, in run_code
    exec(code_obj, self.user_global_ns, self.user_ns)
  File "/tmp/ipykernel_33/4086377508.py", line 1, in <module>
    mlf.fit(
  File "/opt/conda/lib/python3.10/site-packages/mlforecast/utils.py", line 186, in inner
  File "/opt/conda/lib/python3.10/site-packages/mlforecast/forecast.py", line 376, in fit
    X = prep[self.ts.features_order_]
  File "/opt/conda/lib/python3.10/site-packages/mlforecast/forecast.py", line 308, in _conformity_scores
    h: int = 1,
  File "/opt/conda/lib/python3.10/site-packages/mlforecast/utils.py", line 186, in inner
  File "/opt/conda/lib/python3.10/site-packages/mlforecast/forecast.py", line 685, in cross_validation
    keep_last_n=self.ts.keep_last_n,
  File "/opt/conda/lib/python3.10/site-packages/mlforecast/utils.py", line 186, in inner
  File "/opt/conda/lib/python3.10/site-packages/mlforecast/forecast.py", line 489, in predict
    self._cs_df: Optional[DataFrame] = None
  File "/opt/conda/lib/python3.10/site-packages/mlforecast/core.py", line 575, in predict
    self.curr_dates = self.curr_dates[self._idxs]
  File "/opt/conda/lib/python3.10/site-packages/mlforecast/core.py", line 505, in _predict_recursive
    )
UnboundLocalError: local variable 'preds' referenced before assignment

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "/opt/conda/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 2144, in showtraceback
    stb = self.InteractiveTB.structured_traceback(
  File "/opt/conda/lib/python3.10/site-packages/IPython/core/ultratb.py", line 1435, in structured_traceback
    return FormattedTB.structured_traceback(
  File "/opt/conda/lib/python3.10/site-packages/IPython/core/ultratb.py", line 1326, in structured_traceback
    return VerboseTB.structured_traceback(
  File "/opt/conda/lib/python3.10/site-packages/IPython/core/ultratb.py", line 1173, in structured_traceback
    formatted_exception = self.format_exception_as_a_whole(etype, evalue, etb, number_of_lines_of_context,
  File "/opt/conda/lib/python3.10/site-packages/IPython/core/ultratb.py", line 1088, in format_exception_as_a_whole
    frames.append(self.format_record(record))
  File "/opt/conda/lib/python3.10/site-packages/IPython/core/ultratb.py", line 970, in format_record
    frame_info.lines, Colors, self.has_colors, lvals
  File "/opt/conda/lib/python3.10/site-packages/IPython/core/ultratb.py", line 792, in lines
    return self._sd.lines
  File "/opt/conda/lib/python3.10/site-packages/stack_data/utils.py", line 145, in cached_property_wrapper
    value = obj.__dict__[self.func.__name__] = self.func(obj)
  File "/opt/conda/lib/python3.10/site-packages/stack_data/core.py", line 734, in lines
    pieces = self.included_pieces
  File "/opt/conda/lib/python3.10/site-packages/stack_data/utils.py", line 145, in cached_property_wrapper
    value = obj.__dict__[self.func.__name__] = self.func(obj)
  File "/opt/conda/lib/python3.10/site-packages/stack_data/core.py", line 681, in included_pieces
    pos = scope_pieces.index(self.executing_piece)
  File "/opt/conda/lib/python3.10/site-packages/stack_data/utils.py", line 145, in cached_property_wrapper
    value = obj.__dict__[self.func.__name__] = self.func(obj)
  File "/opt/conda/lib/python3.10/site-packages/stack_data/core.py", line 660, in executing_piece
    return only(
  File "/opt/conda/lib/python3.10/site-packages/executing/executing.py", line 116, in only
    raise NotOneValueFound('Expected one value, found 0')
executing.executing.NotOneValueFound: Expected one value, found 0

hello people! I have some trouble using APIs. After I trained the TimesNet Model, what should i do if i want to save the model and predict values? How can I know the datetime of the predicted value

Hello everyone, I hope you are all well. I would like to quickly clarify a doubt about mlforecast. During the transfer learning process

model_lgbm.predict(h=15, new_df=new_skus_df)

Is it possible to parallelize this process? Because there are more than 300k distinct series Thanks in advance!

i upgraded my neuralforecast pip module version and im getting this error when I attempt to run

.predict()

AttributeError: 'TimeSeriesDataset' object has no attribute 'scalers_'

Hi all, looking for any recommendarions and suggestions for limiting model forecast change coming off of seasonal periods. I have a global ML tree model (lgb, xbf, catboost) at the weekly level which correctly models seasonality patterns, but when i run a model right after the seasonal spike, vs 2 months later when in the seasonal valley, the Level drastically changes. My assumption is due to the recent lags are high after the spike, and low during the valley What happens is the long term volume changes by 30% or more just bc of the last 2 months of data. I already use fourier features for week and month I do 4 week recursive forecast with lags starting at 4 Weekly forecasts with 2 year horizon I do lags, rolling MA, diffs I have tried doing longer recursive windows of 13 and 26. I also tried adjusting l1 and l2 Impacts boosted algos most, random forecast isn't impacted as much I have tried de-weighting the most recent values, which actually worked but there's no justification to de-weight your most recent data Curious if others have this issue

Hello all, I will like to know if anyone has any experience working with multiple time series, whereby you pass each series through some tests and the result of the tests provides an indication that the series belongs to a certain group (something like time series clustering but without training a model). I have a panel data set containing a lot of series which are different in time periods and scale. I want to separate all of them into different groups and apply the appropriate preprocessing steps to them before performing forecasting. TIA!

Hi everyone. I have a general question (I believe more in the computational side, less of the modelling aspect). Imagine a scenario that I have a timeseries (with exogenous covariates), and I am able to successfully train a model (given a set of metrics) for forecasting a given horizon. How much of a problem would it be using that same model for “backcasting”, in complement of forecasting? For simplicity, I consider that conditions of the model training to be the same in previous time (when compared to the timeseries sample used for training).

There have been prizes announced for the probabilistic forecasting competition, including free enrolment into “Applied Conformal Prediction’ live instructor-led course for the best solution utilising conformal prediction 🏆🔥🚀https://www.kaggle.com/competitions/probabilistic-forecasting-i-temperature/discussion/493217

Daylight savings & Time-index Hi everyone, I was just wondering how one might go about creating a five-minutely time index that is based on the clock-datetime (so it accounts for daylight savings) instead of the UTC time (what I have in my

ds

column)? I've used the function below, which works fine to create the five-minutely time index based on the UTC time (

ds

column):

def five_min_index(dates):
    """Calculate 5-minutely index for each datetime (0 to 287)"""
    return (dates.hour * 60 + dates.minute) // 5

It is used in

LightGBMCV()

as follows:

cv = LightGBMCV(
    freq='5min',
    target_transforms=[Differences([288])],
    lags=[1,2,3,4,5,6,12,288],
    lag_transforms={
        1: [
            ExponentiallyWeightedMean(alpha=0.5), 
            RollingMean(window_size=12), 
        ],
        12: [RollingMean(window_size=288)],
    },
    date_features=[five_min_index, 'hour', 'dayofweek'],
    num_threads=4,
)

My

y

depends on the clock-datetime, it's human driven, so is influenced by when we start the day and when we end the day. Any thoughts? I'd like to be able to have the index in

date_features

instead of having to create a dynamic exogenous feature...

Hello Team, I have a question regarding to the model training associated with the 'id_col' in the data. And the 'id_col' exits in NeuralForecast, MLForecast, and TimeGPT As we know that 'id_col' is symbol for multi-series. And for the model that naturally can have multi-series as input (partly in NeuralForecast), the trained model will be the global model (one model for all). However, for the model that does not naturally have multi-series as input (most in MLForecast), I wonder it will have three possible scenarios. 1. Automatic add 'id_col' to 'static_features', with the ordinal encoding (categorial value to integer), so the 'id_col' becomes an extra features to training the global model. 2. Training individual models for each series ('id') in the multi-series, and so the resultant trained models would be costly if the multi-series has many unique series. 3. Training the global model for all the series, ignore the 'id_col' (drop it), and so the resultant trained model would output same value for every series (the series has no intrinsic difference). Please correctify me and explain what's the actual strategy when a model (which naturally does not handle multi-series) dealing with multi-series that we use for training. Many Thanks!

hello! I have a problem, if I have a time sequence whose length is 10. is it possible that i use the model to forecast data? Here are my api:

horizon = 2

models = [NHITS(h=horizon,
               input_size=2*horizon,
               max_steps=50),
         NBEATS(h=horizon,
               input_size=2*horizon,
               max_steps=50),
         TimesNet(h=horizon,
                 input_size=2*horizon,
                 max_steps=50)]

#%%
nf = NeuralForecast(models=models, freq='H')
#%%
a=nf.cross_validation(df=data, step_size=horizon, n_windows=1)

By the way, i wanna know what is mas_Steps, is this iteration times? and what is input_size?

Hey, I'd like to thank you for developing and maintaining such a precious library. Not only this but also SHOUT-OUT to the documenting in a really beginner friendly way. I will be using the library for my master thesis on a time series + transfer learning application, and therefore would like to send my gratitude!

Hi, I'm using NHITS to identify outliers/anomalies in time series data. My time series data is hierarchical. Are there any useful libraries that can help perform RCA (root cause analysis) on the identified anomalies? Based on the RCA, it'd be easier to then recommend data correction based on business logics.

Hello everyone, I hope you're all doing well! Just a quick note to say thanks for the awesome library (the machine learning engineers were really giving me a hard time for using R to forecasting : / ), Nixtlaverse it s realy awesome! And also to share the results of the global model I've been working on over the past year. About the company I work for, Olist – think of us as the bridge connecting sellers to buyers across multiple marketplaces in Brazil. We’re all about that e-commerce life, and we've got a mountain of products (over a million SKUs) to keep an eye on. We faced some critical challenges in finding valid historical data in terms of both sales quantity and variation, since a large part of our products are long-tail; that is, most sell once and stop, others sell occasionally but when they do, it’s just a single unit, and some sold well in the past but have stopped for months... etc. After carefully selecting our training data and meticulously engineering the variables, I trained a global LightGBM focusing on maximizing overall R2 (so the model can handle different magnitudes of sales) and concurrently evaluating individual WMAPE. Once the model was trained, I applied transfer learning to nearly 70k SKUs in addition to generating confidence intervals via bootstrap. The results were evaluated by types of demand: Here’s the scoop on our forecast accuracy (median WMAPE by demand pattern): Erratic demand: 36% - Intermittent demand: 43% - Lumpy demand: 55% - Smooth demand: 27% And for confidence interval coverage, intermittent demand took the lead with nearly 80% coverage! Some plots and metrics in the attached image.